chhatramani/nepali-fake-news-dataset-v1

📊 Nepali Fake News Dataset (Synthetic & Balanced)

📌 Overview

This dataset contains 8,838 labeled Nepali news samples designed for Fake News Detection tasks.
Due to the scarcity of labeled Nepali fake news datasets, this corpus was synthetically generated using Large Language Models (OpenAI GPT & Google Gemini).
It covers diverse topics relevant to Nepal, including politics, entertainment, infrastructure, health, economy, agriculture, and technology.

The dataset is perfectly balanced with 50% REAL and 50% FAKE samples, making it ideal for training and evaluating binary text classification models.

---

📊 Dataset Statistics

• Total Samples: 8,838
• Classes: 2 (REAL, FAKE)
• Balance: Perfectly Balanced (4,419 REAL / 4,419 FAKE)
• Language: Nepali (Devanagari Script)
• Source: Synthetic Generation (OpenAI GPT & Google Gemini)

Label	Count	Percentage
REAL	4,419	50%
FAKE	4,419	50%

---

Columns Description

• label: Target variable (REAL or FAKE).
• category: Domain of the news (e.g., Politics & Government, Entertainment & Cinema, Economy & Business).
• headline: Title of the news article (short text).
• body: Main content of the news article (medium to long text).
• full_text: Concatenated version of headline + body, ready for direct tokenization in models like BERT or XLM-RoBERTa.

---

🎯 Key Features

• Diversity: Covers 8+ categories including Politics, Sports, Agriculture, Technology, and more.
• Stylistic Variation:
  • FAKE samples mimic misinformation traits (excessive punctuation, emotional language, vague sourcing).
  • REAL samples follow formal journalistic styles.
• Pre-processed: Already shuffled, balanced, and cleaned (normalized Unicode).
• Ready-to-use: Suitable for immediate training without additional preprocessing.

---

💡 Use Cases

This dataset is ideal for:

• Fine-tuning Transformer Models: Train models like XLM-RoBERTa, mBERT, or DistilBERT for Nepali fake news detection.
• Text Classification Benchmarks: Evaluate algorithms on a low-resource language with complex morphology.
• Stylistic Analysis: Study linguistic differences between formal journalism and sensationalist rumor-mongering.
• Cross-lingual Transfer Learning: Use Nepali as a case study for transferring knowledge from high-resource languages.
• Explainable AI (XAI): Apply interpretability methods (e.g., SHAP, LIME) to understand features driving fake vs. real classification.

---

🤖 Data Generation Methodology

The dataset was created using a synthetic pipeline:

• OpenAI GPT models: Generated context-aware, high-quality text.
• Google Gemini models: Produced diverse, topic-specific content in Nepali.
• Prompt Engineering: Controlled stylistic differences (formal vs. sensational).

---

⚠️ Important Notes

• Synthetic Content: The dataset is fictional. Events, people, and statistics may not reflect real-world facts.
• Research Purpose Only: Intended for machine learning research and text classification tasks.
• AI Bias: May inherit biases from generative models or prompts.
• Real-world Gap: Performance on synthetic data may not directly translate to real-world Nepali social media or news.

---

🛠️ Suggested Workflow

1. Load the CSV file using Pandas.
2. Split into Train/Test/Validation sets (e.g., 80/10/10).
3. Use the full_text column for model input.
4. Tokenize using a multilingual tokenizer (e.g., sentencepiece for XLM-RoBERTa).
5. Train a binary classifier.

---

📝 Citation

If you use this dataset in your research, please cite:

@dataset{yadav_nepali_fake_news_2026,
  title     = {Nepali Fake News Dataset V1},
  author    = {Chhatramani Yadav},
  year      = {2026},
  publisher = {Hugging Face Datasets},
  url       = {https://huggingface.co/datasets/chhatramani/nepali-fake-news-dataset-v1/}
}